NeuroFarm

Closed-loop crop control · Neuromorphic edge AI · Amsterdam

The offline brain for agriculture.

NeuroFarm controls irrigation, lighting and nutrients with a single spiking neural network — trained with reinforcement learning, running on neuromorphic hardware right next to the plant. No cloud. No hand-written rules. Every actuator command is one forward pass.

Request the investor brief See how it works
Sensors → actuators
5 → 3
Decisions / day
288
Parameters
7,259
Power budget
< 1 W design target
Fig. 1 — The closed loop Illustrative
Decision 073 / 288 · 06:05
Illustrative animation, not model output: spike timings, network activity and actuator levels are generated to explain the idea, not recorded from NeuroFarm’s policy or hardware. The four scenarios are the ones the real policy is trained and tested under.
Status

Phase 1 in progress: moving onto real neuromorphic hardware at the VU Amsterdam Demonstrator Lab. Research release v0.1.0 shipped May 2026. No measured performance results yet — energy will be published as ranges measured on real silicon.

01 — The problem

Greenhouses are full of sensors. Their control is still hand-written rules.

1.1

Control that can’t adapt

Hand-tuned thresholds, fixed LED schedules and PID loops keep control conservative. They don’t learn from the crop in front of them.

1.2

A rising footprint

Indoor and vertical farming carry a rising water and energy footprint for every kilogram of produce.

1.3

Cloud AI, bolted on top

Most “AI in agriculture” runs in the cloud, on top of those same rules — adding latency, bandwidth and energy cost, and failing where connectivity is poor: rural sites and open fields.

Today
Sensors Internet uplink Cloud model Thresholds · PID · timers Actuators
NeuroFarm
5 sensors One spiking network, on-site, on a neuromorphic chip 3 actuators No internet · no rules in the loop

02 — How it works

From a scientific digital twin to a policy that fits in on-chip memory.

Five sensors in, three actuators out, one decision every five minutes — 288 a day. The control policy is learned in simulation, compressed, then converted into a spiking neural network that runs on a BrainChip Akida AKD1000 beside the crop.

  1. Simulate

    A scientific lettuce-growth digital twin (Van Henten model) plus a climate model. 47 crop parameters, each traced to a peer-reviewed source.

  2. Learn

    A reinforcement-learning agent (Soft Actor-Critic) learns the control policy inside the twin.

  3. Distil

    The policy is distilled into a tiny network: 5 → 32 → 16 → 3.

  4. Quantise

    Quantised to 8/4/4-bit so the whole policy fits in on-chip memory.

  5. Deploy

    Converted into a spiking neural network running on a BrainChip Akida AKD1000 — right next to the plant.

Network profile

parameters
7,259
MACs per decision
8,760
activation sparsity
~50%
bit quantisation
8/4/4
layer widths
5·32·16·3
whole policy in on-chip memory
On-chip

The closed loop

5 sensor inputs

  • Soil moisture
  • Air temperature / humidity
  • Light (PPFD)
  • CO₂
  • Soil electrical conductivity (EC)

3 actuator outputs

  • Irrigation pump
  • LED intensity
  • Nutrient dosing pump

Phase 1 edge node

Akida AKD1000 + Raspberry Pi 5 + inline power metering. No thresholds, if-else rules or PID loops in the control path.

Trained & tested under NominalDroughtOver-irrigationCO₂ drop
Get the technical brief →

03 — Scientific rigour

We wrote down what would count as success before running a single benchmark.

Nine hypotheses were pre-registered in version control on 10 May 2026, before any benchmark data existed. Every headline comparison is made against the controllers growers use today — and energy is measured on real silicon, reported as honest ranges.

Pre-registration Committed to version control
Date
10 May 2026
Hypotheses
9
Benchmark data
None existed yet
Baselines
PID · rule-based · timer schedules
Statistics
Wilcoxon · Cliff’s δ · Holm–Bonferroni
pre-registered hypotheses
9
benchmark runs
210
seeds per headline comparison
20
automated tests
366
architecture decision records
25+
research release, May 2026
v0.1.0

What we don’t claim yet: there are no measured performance results. Water, energy or yield savings will only appear here once they’ve been measured and tested against the pre-registered hypotheses.

04 — Traction & roadmap

Phase 0 is done. Now we’re moving onto real silicon.

  1. Oct 2025 – Jun 2026

    Phase 0 Complete

    Digital twin, full RL → SNN toolchain, pre-registered benchmark protocol.

  2. Feb 2026

    Accepted into VU D-Lab

    The VU Amsterdam Demonstrator Lab, a European deep-tech program.

  3. 2026 · We are here

    Phase 1 In progress

    Running on real neuromorphic hardware at VU D-Lab. Silicon-level power measurement. Peer-reviewed publication.

  4. Next

    Phase 1.5

    A camera with spiking vision on the same chip: disease detection, biomass estimation, visual feedback into control.

  5. 2027+

    Phase 2

    Multiple crops and cultivars, field pilots with partners, LED-spectrum control.

Long-term vision

Vertical farmGreenhouseOpen fieldSpace agriculture

05 — Market

Start indoors. Then greenhouses. Then the open field.

  1. Indoor & vertical farms

    Our beachhead.

  2. Commercial greenhouses

    An alternative to rule-based climate computers.

  3. Open field

    Where working offline matters most.

Conceptual — not to scale

Tailwinds

  • EU Green Deal pressure on water, fertiliser and energy use
  • Food security in climate-stressed regions
  • Falling cost of neuromorphic and edge AI chips
  • Strong European focus on deep-tech AgriTech

06 — Team

One founder on the silicon. One on the market.

Accepted into the VU Amsterdam Demonstrator Lab (D-Lab) in February 2026.

07 — Contact

Request the investor brief.

For pre-seed and seed investors, greenhouse and vertical-farm operators interested in a pilot, and research partners. We reply personally.

Web
neurofarm.nl
Based at
VU Amsterdam Demonstrator Lab
Amsterdam, Netherlands

We only use your details to reply to you.